Use per-mount kmalloc pools for bulk data structures, particularly inodes
[dragonfly.git] / sys / vfs / hammer / hammer_flusher.c
1 /*
2  * Copyright (c) 2008 The DragonFly Project.  All rights reserved.
3  * 
4  * This code is derived from software contributed to The DragonFly Project
5  * by Matthew Dillon <dillon@backplane.com>
6  * 
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in
15  *    the documentation and/or other materials provided with the
16  *    distribution.
17  * 3. Neither the name of The DragonFly Project nor the names of its
18  *    contributors may be used to endorse or promote products derived
19  *    from this software without specific, prior written permission.
20  * 
21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  * 
34  * $DragonFly: src/sys/vfs/hammer/hammer_flusher.c,v 1.45 2008/07/31 04:42:04 dillon Exp $
35  */
36 /*
37  * HAMMER dependancy flusher thread
38  *
39  * Meta data updates create buffer dependancies which are arranged as a
40  * hierarchy of lists.
41  */
42
43 #include "hammer.h"
44
45 static void hammer_flusher_master_thread(void *arg);
46 static void hammer_flusher_slave_thread(void *arg);
47 static void hammer_flusher_flush(hammer_mount_t hmp);
48 static void hammer_flusher_flush_inode(hammer_inode_t ip,
49                                         hammer_transaction_t trans);
50
51 /*
52  * Support structures for the flusher threads.
53  */
54 struct hammer_flusher_info {
55         TAILQ_ENTRY(hammer_flusher_info) entry;
56         struct hammer_mount *hmp;
57         thread_t        td;
58         int             runstate;
59         int             count;
60         hammer_flush_group_t flg;
61         hammer_inode_t  work_array[HAMMER_FLUSH_GROUP_SIZE];
62 };
63
64 typedef struct hammer_flusher_info *hammer_flusher_info_t;
65
66 /*
67  * Sync all inodes pending on the flusher.
68  *
69  * All flush groups will be flushed.  This does not queue dirty inodes
70  * to the flush groups, it just flushes out what has already been queued!
71  */
72 void
73 hammer_flusher_sync(hammer_mount_t hmp)
74 {
75         int seq;
76
77         seq = hammer_flusher_async(hmp, NULL);
78         hammer_flusher_wait(hmp, seq);
79 }
80
81 /*
82  * Sync all inodes pending on the flusher - return immediately.
83  *
84  * All flush groups will be flushed.
85  */
86 int
87 hammer_flusher_async(hammer_mount_t hmp, hammer_flush_group_t close_flg)
88 {
89         hammer_flush_group_t flg;
90         int seq = hmp->flusher.next;
91
92         TAILQ_FOREACH(flg, &hmp->flush_group_list, flush_entry) {
93                 if (flg->running == 0)
94                         ++seq;
95                 flg->closed = 1;
96                 if (flg == close_flg)
97                         break;
98         }
99         if (hmp->flusher.td) {
100                 if (hmp->flusher.signal++ == 0)
101                         wakeup(&hmp->flusher.signal);
102         } else {
103                 seq = hmp->flusher.done;
104         }
105         return(seq);
106 }
107
108 int
109 hammer_flusher_async_one(hammer_mount_t hmp)
110 {
111         int seq;
112
113         if (hmp->flusher.td) {
114                 seq = hmp->flusher.next;
115                 if (hmp->flusher.signal++ == 0)
116                         wakeup(&hmp->flusher.signal);
117         } else {
118                 seq = hmp->flusher.done;
119         }
120         return(seq);
121 }
122
123 /*
124  * Wait for the flusher to get to the specified sequence number.
125  * Signal the flusher as often as necessary to keep it going.
126  */
127 void
128 hammer_flusher_wait(hammer_mount_t hmp, int seq)
129 {
130         while ((int)(seq - hmp->flusher.done) > 0) {
131                 if (hmp->flusher.act != seq) {
132                         if (hmp->flusher.signal++ == 0)
133                                 wakeup(&hmp->flusher.signal);
134                 }
135                 tsleep(&hmp->flusher.done, 0, "hmrfls", 0);
136         }
137 }
138
139 void
140 hammer_flusher_create(hammer_mount_t hmp)
141 {
142         hammer_flusher_info_t info;
143         int i;
144
145         hmp->flusher.signal = 0;
146         hmp->flusher.act = 0;
147         hmp->flusher.done = 0;
148         hmp->flusher.next = 1;
149         hammer_ref(&hmp->flusher.finalize_lock);
150         TAILQ_INIT(&hmp->flusher.run_list);
151         TAILQ_INIT(&hmp->flusher.ready_list);
152
153         lwkt_create(hammer_flusher_master_thread, hmp,
154                     &hmp->flusher.td, NULL, 0, -1, "hammer-M");
155         for (i = 0; i < HAMMER_MAX_FLUSHERS; ++i) {
156                 info = kmalloc(sizeof(*info), hmp->m_misc, M_WAITOK|M_ZERO);
157                 info->hmp = hmp;
158                 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry);
159                 lwkt_create(hammer_flusher_slave_thread, info,
160                             &info->td, NULL, 0, -1, "hammer-S%d", i);
161         }
162 }
163
164 void
165 hammer_flusher_destroy(hammer_mount_t hmp)
166 {
167         hammer_flusher_info_t info;
168
169         /*
170          * Kill the master
171          */
172         hmp->flusher.exiting = 1;
173         while (hmp->flusher.td) {
174                 ++hmp->flusher.signal;
175                 wakeup(&hmp->flusher.signal);
176                 tsleep(&hmp->flusher.exiting, 0, "hmrwex", hz);
177         }
178
179         /*
180          * Kill the slaves
181          */
182         while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) != NULL) {
183                 KKASSERT(info->runstate == 0);
184                 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
185                 info->runstate = -1;
186                 wakeup(&info->runstate);
187                 while (info->td)
188                         tsleep(&info->td, 0, "hmrwwc", 0);
189                 TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
190                 kfree(info, hmp->m_misc);
191         }
192 }
193
194 /*
195  * The master flusher thread manages the flusher sequence id and
196  * synchronization with the slave work threads.
197  */
198 static void
199 hammer_flusher_master_thread(void *arg)
200 {
201         hammer_flush_group_t flg;
202         hammer_mount_t hmp;
203
204         hmp = arg;
205
206         for (;;) {
207                 /*
208                  * Do at least one flush cycle.  We may have to update the
209                  * UNDO FIFO even if no inodes are queued.
210                  */
211                 for (;;) {
212                         while (hmp->flusher.group_lock)
213                                 tsleep(&hmp->flusher.group_lock, 0, "hmrhld", 0);
214                         hmp->flusher.act = hmp->flusher.next;
215                         ++hmp->flusher.next;
216                         hammer_flusher_clean_loose_ios(hmp);
217                         hammer_flusher_flush(hmp);
218                         hmp->flusher.done = hmp->flusher.act;
219                         wakeup(&hmp->flusher.done);
220                         flg = TAILQ_FIRST(&hmp->flush_group_list);
221                         if (flg == NULL || flg->closed == 0)
222                                 break;
223                         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
224                                 break;
225                 }
226
227                 /*
228                  * Wait for activity.
229                  */
230                 if (hmp->flusher.exiting && TAILQ_EMPTY(&hmp->flush_group_list))
231                         break;
232                 while (hmp->flusher.signal == 0)
233                         tsleep(&hmp->flusher.signal, 0, "hmrwwa", 0);
234
235                 /*
236                  * Flush for each count on signal but only allow one extra
237                  * flush request to build up.
238                  */
239                 if (--hmp->flusher.signal != 0)
240                         hmp->flusher.signal = 1;
241         }
242
243         /*
244          * And we are done.
245          */
246         hmp->flusher.td = NULL;
247         wakeup(&hmp->flusher.exiting);
248         lwkt_exit();
249 }
250
251 /*
252  * Flush all inodes in the current flush group.
253  */
254 static void
255 hammer_flusher_flush(hammer_mount_t hmp)
256 {
257         hammer_flusher_info_t info;
258         hammer_flush_group_t flg;
259         hammer_reserve_t resv;
260         hammer_inode_t ip;
261         hammer_inode_t next_ip;
262         int slave_index;
263         int count;
264
265         /*
266          * Just in-case there's a flush race on mount
267          */
268         if (TAILQ_FIRST(&hmp->flusher.ready_list) == NULL)
269                 return;
270
271         /*
272          * We only do one flg but we may have to loop/retry.
273          */
274         count = 0;
275         while ((flg = TAILQ_FIRST(&hmp->flush_group_list)) != NULL) {
276                 ++count;
277                 if (hammer_debug_general & 0x0001) {
278                         kprintf("hammer_flush %d ttl=%d recs=%d\n",
279                                 hmp->flusher.act,
280                                 flg->total_count, flg->refs);
281                 }
282                 if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
283                         break;
284                 hammer_start_transaction_fls(&hmp->flusher.trans, hmp);
285
286                 /*
287                  * If the previous flush cycle just about exhausted our
288                  * UNDO space we may have to do a dummy cycle to move the
289                  * first_offset up before actually digging into a new cycle,
290                  * or the new cycle will not have sufficient undo space.
291                  */
292                 if (hammer_flusher_undo_exhausted(&hmp->flusher.trans, 3))
293                         hammer_flusher_finalize(&hmp->flusher.trans, 0);
294
295                 /*
296                  * Ok, we are running this flush group now (this prevents new
297                  * additions to it).
298                  */
299                 flg->running = 1;
300                 if (hmp->next_flush_group == flg)
301                         hmp->next_flush_group = TAILQ_NEXT(flg, flush_entry);
302
303                 /*
304                  * Iterate the inodes in the flg's flush_list and assign
305                  * them to slaves.
306                  */
307                 slave_index = 0;
308                 info = TAILQ_FIRST(&hmp->flusher.ready_list);
309                 next_ip = TAILQ_FIRST(&flg->flush_list);
310
311                 while ((ip = next_ip) != NULL) {
312                         next_ip = TAILQ_NEXT(ip, flush_entry);
313
314                         /*
315                          * Add ip to the slave's work array.  The slave is
316                          * not currently running.
317                          */
318                         info->work_array[info->count++] = ip;
319                         if (info->count != HAMMER_FLUSH_GROUP_SIZE)
320                                 continue;
321
322                         /*
323                          * Get the slave running
324                          */
325                         TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
326                         TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry);
327                         info->flg = flg;
328                         info->runstate = 1;
329                         wakeup(&info->runstate);
330
331                         /*
332                          * Get a new slave.  We may have to wait for one to
333                          * finish running.
334                          */
335                         while ((info = TAILQ_FIRST(&hmp->flusher.ready_list)) == NULL) {
336                                 tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0);
337                         }
338                 }
339
340                 /*
341                  * Run the current slave if necessary
342                  */
343                 if (info->count) {
344                         TAILQ_REMOVE(&hmp->flusher.ready_list, info, entry);
345                         TAILQ_INSERT_TAIL(&hmp->flusher.run_list, info, entry);
346                         info->flg = flg;
347                         info->runstate = 1;
348                         wakeup(&info->runstate);
349                 }
350
351                 /*
352                  * Wait for all slaves to finish running
353                  */
354                 while (TAILQ_FIRST(&hmp->flusher.run_list) != NULL)
355                         tsleep(&hmp->flusher.ready_list, 0, "hmrfcc", 0);
356
357                 /*
358                  * Do the final finalization, clean up
359                  */
360                 hammer_flusher_finalize(&hmp->flusher.trans, 1);
361                 hmp->flusher.tid = hmp->flusher.trans.tid;
362
363                 hammer_done_transaction(&hmp->flusher.trans);
364
365                 /*
366                  * Loop up on the same flg.  If the flg is done clean it up
367                  * and break out.  We only flush one flg.
368                  */
369                 if (TAILQ_FIRST(&flg->flush_list) == NULL) {
370                         KKASSERT(TAILQ_EMPTY(&flg->flush_list));
371                         KKASSERT(flg->refs == 0);
372                         TAILQ_REMOVE(&hmp->flush_group_list, flg, flush_entry);
373                         kfree(flg, hmp->m_misc);
374                         break;
375                 }
376         }
377
378         /*
379          * We may have pure meta-data to flush, or we may have to finish
380          * cycling the UNDO FIFO, even if there were no flush groups.
381          */
382         if (count == 0 && hammer_flusher_haswork(hmp)) {
383                 hammer_start_transaction_fls(&hmp->flusher.trans, hmp);
384                 hammer_flusher_finalize(&hmp->flusher.trans, 1);
385                 hammer_done_transaction(&hmp->flusher.trans);
386         }
387
388         /*
389          * Clean up any freed big-blocks (typically zone-2). 
390          * resv->flush_group is typically set several flush groups ahead
391          * of the free to ensure that the freed block is not reused until
392          * it can no longer be reused.
393          */
394         while ((resv = TAILQ_FIRST(&hmp->delay_list)) != NULL) {
395                 if (resv->flush_group != hmp->flusher.act)
396                         break;
397                 hammer_reserve_clrdelay(hmp, resv);
398         }
399 }
400
401
402 /*
403  * The slave flusher thread pulls work off the master flush_list until no
404  * work is left.
405  */
406 static void
407 hammer_flusher_slave_thread(void *arg)
408 {
409         hammer_flush_group_t flg;
410         hammer_flusher_info_t info;
411         hammer_mount_t hmp;
412         hammer_inode_t ip;
413         int i;
414
415         info = arg;
416         hmp = info->hmp;
417
418         for (;;) {
419                 while (info->runstate == 0)
420                         tsleep(&info->runstate, 0, "hmrssw", 0);
421                 if (info->runstate < 0)
422                         break;
423                 flg = info->flg;
424
425                 for (i = 0; i < info->count; ++i) {
426                         ip = info->work_array[i];
427                         hammer_flusher_flush_inode(ip, &hmp->flusher.trans);
428                         ++hammer_stats_inode_flushes;
429                 }
430                 info->count = 0;
431                 info->runstate = 0;
432                 TAILQ_REMOVE(&hmp->flusher.run_list, info, entry);
433                 TAILQ_INSERT_TAIL(&hmp->flusher.ready_list, info, entry);
434                 wakeup(&hmp->flusher.ready_list);
435         }
436         info->td = NULL;
437         wakeup(&info->td);
438         lwkt_exit();
439 }
440
441 void
442 hammer_flusher_clean_loose_ios(hammer_mount_t hmp)
443 {
444         hammer_buffer_t buffer;
445         hammer_io_t io;
446
447         /*
448          * loose ends - buffers without bp's aren't tracked by the kernel
449          * and can build up, so clean them out.  This can occur when an
450          * IO completes on a buffer with no references left.
451          */
452         if ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) {
453                 crit_enter();   /* biodone() race */
454                 while ((io = TAILQ_FIRST(&hmp->lose_list)) != NULL) {
455                         KKASSERT(io->mod_list == &hmp->lose_list);
456                         TAILQ_REMOVE(&hmp->lose_list, io, mod_entry);
457                         io->mod_list = NULL;
458                         if (io->lock.refs == 0)
459                                 ++hammer_count_refedbufs;
460                         hammer_ref(&io->lock);
461                         buffer = (void *)io;
462                         hammer_rel_buffer(buffer, 0);
463                 }
464                 crit_exit();
465         }
466 }
467
468 /*
469  * Flush a single inode that is part of a flush group.
470  *
471  * Flusher errors are extremely serious, even ENOSPC shouldn't occur because
472  * the front-end should have reserved sufficient space on the media.  Any
473  * error other then EWOULDBLOCK will force the mount to be read-only.
474  */
475 static
476 void
477 hammer_flusher_flush_inode(hammer_inode_t ip, hammer_transaction_t trans)
478 {
479         hammer_mount_t hmp = ip->hmp;
480         int error;
481
482         hammer_flusher_clean_loose_ios(hmp);
483         error = hammer_sync_inode(trans, ip);
484
485         /*
486          * EWOULDBLOCK can happen under normal operation, all other errors
487          * are considered extremely serious.  We must set WOULDBLOCK
488          * mechanics to deal with the mess left over from the abort of the
489          * previous flush.
490          */
491         if (error) {
492                 ip->flags |= HAMMER_INODE_WOULDBLOCK;
493                 if (error == EWOULDBLOCK)
494                         error = 0;
495         }
496         hammer_flush_inode_done(ip, error);
497         while (hmp->flusher.finalize_want)
498                 tsleep(&hmp->flusher.finalize_want, 0, "hmrsxx", 0);
499         if (hammer_flusher_undo_exhausted(trans, 1)) {
500                 kprintf("HAMMER: Warning: UNDO area too small!\n");
501                 hammer_flusher_finalize(trans, 1);
502         } else if (hammer_flusher_meta_limit(trans->hmp)) {
503                 hammer_flusher_finalize(trans, 0);
504         }
505 }
506
507 /*
508  * Return non-zero if the UNDO area has less then (QUARTER / 4) of its
509  * space left.
510  *
511  * 1/4 - Emergency free undo space level.  Below this point the flusher
512  *       will finalize even if directory dependancies have not been resolved.
513  *
514  * 2/4 - Used by the pruning and reblocking code.  These functions may be
515  *       running in parallel with a flush and cannot be allowed to drop
516  *       available undo space to emergency levels.
517  *
518  * 3/4 - Used at the beginning of a flush to force-sync the volume header
519  *       to give the flush plenty of runway to work in.
520  */
521 int
522 hammer_flusher_undo_exhausted(hammer_transaction_t trans, int quarter)
523 {
524         if (hammer_undo_space(trans) <
525             hammer_undo_max(trans->hmp) * quarter / 4) {
526                 return(1);
527         } else {
528                 return(0);
529         }
530 }
531
532 /*
533  * Flush all pending UNDOs, wait for write completion, update the volume
534  * header with the new UNDO end position, and flush it.  Then
535  * asynchronously flush the meta-data.
536  *
537  * If this is the last finalization in a flush group we also synchronize
538  * our cached blockmap and set hmp->flusher_undo_start and our cached undo
539  * fifo first_offset so the next flush resets the FIFO pointers.
540  *
541  * If this is not final it is being called because too many dirty meta-data
542  * buffers have built up and must be flushed with UNDO synchronization to
543  * avoid a buffer cache deadlock.
544  */
545 void
546 hammer_flusher_finalize(hammer_transaction_t trans, int final)
547 {
548         hammer_volume_t root_volume;
549         hammer_blockmap_t cundomap, dundomap;
550         hammer_mount_t hmp;
551         hammer_io_t io;
552         int count;
553         int i;
554
555         hmp = trans->hmp;
556         root_volume = trans->rootvol;
557
558         /*
559          * Exclusively lock the flusher.  This guarantees that all dirty
560          * buffers will be idled (have a mod-count of 0).
561          */
562         ++hmp->flusher.finalize_want;
563         hammer_lock_ex(&hmp->flusher.finalize_lock);
564
565         /*
566          * If this isn't the final sync several threads may have hit the
567          * meta-limit at the same time and raced.  Only sync if we really
568          * have to, after acquiring the lock.
569          */
570         if (final == 0 && !hammer_flusher_meta_limit(hmp))
571                 goto done;
572
573         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
574                 goto done;
575
576         /*
577          * Flush data buffers.  This can occur asynchronously and at any
578          * time.  We must interlock against the frontend direct-data write
579          * but do not have to acquire the sync-lock yet.
580          */
581         count = 0;
582         while ((io = TAILQ_FIRST(&hmp->data_list)) != NULL) {
583                 if (io->ioerror)
584                         break;
585                 if (io->lock.refs == 0)
586                         ++hammer_count_refedbufs;
587                 hammer_ref(&io->lock);
588                 hammer_io_write_interlock(io);
589                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
590                 hammer_io_flush(io);
591                 hammer_io_done_interlock(io);
592                 hammer_rel_buffer((hammer_buffer_t)io, 0);
593                 ++count;
594         }
595
596         /*
597          * The sync-lock is required for the remaining sequence.  This lock
598          * prevents meta-data from being modified.
599          */
600         hammer_sync_lock_ex(trans);
601
602         /*
603          * If we have been asked to finalize the volume header sync the
604          * cached blockmap to the on-disk blockmap.  Generate an UNDO
605          * record for the update.
606          */
607         if (final) {
608                 cundomap = &hmp->blockmap[0];
609                 dundomap = &root_volume->ondisk->vol0_blockmap[0];
610                 if (root_volume->io.modified) {
611                         hammer_modify_volume(trans, root_volume,
612                                              dundomap, sizeof(hmp->blockmap));
613                         for (i = 0; i < HAMMER_MAX_ZONES; ++i)
614                                 hammer_crc_set_blockmap(&cundomap[i]);
615                         bcopy(cundomap, dundomap, sizeof(hmp->blockmap));
616                         hammer_modify_volume_done(root_volume);
617                 }
618         }
619
620         /*
621          * Flush UNDOs
622          */
623         count = 0;
624         while ((io = TAILQ_FIRST(&hmp->undo_list)) != NULL) {
625                 if (io->ioerror)
626                         break;
627                 KKASSERT(io->modify_refs == 0);
628                 if (io->lock.refs == 0)
629                         ++hammer_count_refedbufs;
630                 hammer_ref(&io->lock);
631                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
632                 hammer_io_flush(io);
633                 hammer_rel_buffer((hammer_buffer_t)io, 0);
634                 ++count;
635         }
636
637         /*
638          * Wait for I/Os to complete
639          */
640         hammer_flusher_clean_loose_ios(hmp);
641         hammer_io_wait_all(hmp, "hmrfl1");
642
643         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
644                 goto failed;
645
646         /*
647          * Update the on-disk volume header with new UNDO FIFO end position
648          * (do not generate new UNDO records for this change).  We have to
649          * do this for the UNDO FIFO whether (final) is set or not.
650          *
651          * Also update the on-disk next_tid field.  This does not require
652          * an UNDO.  However, because our TID is generated before we get
653          * the sync lock another sync may have beat us to the punch.
654          *
655          * This also has the side effect of updating first_offset based on
656          * a prior finalization when the first finalization of the next flush
657          * cycle occurs, removing any undo info from the prior finalization
658          * from consideration.
659          *
660          * The volume header will be flushed out synchronously.
661          */
662         dundomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
663         cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
664
665         if (dundomap->first_offset != cundomap->first_offset ||
666                    dundomap->next_offset != cundomap->next_offset) {
667                 hammer_modify_volume(NULL, root_volume, NULL, 0);
668                 dundomap->first_offset = cundomap->first_offset;
669                 dundomap->next_offset = cundomap->next_offset;
670                 hammer_crc_set_blockmap(dundomap);
671                 hammer_modify_volume_done(root_volume);
672         }
673
674         /*
675          * vol0_next_tid is used for TID selection and is updated without
676          * an UNDO so we do not reuse a TID that may have been rolled-back.
677          *
678          * vol0_last_tid is the highest fully-synchronized TID.  It is
679          * set-up when the UNDO fifo is fully synced, later on (not here).
680          */
681         if (root_volume->io.modified) {
682                 hammer_modify_volume(NULL, root_volume, NULL, 0);
683                 if (root_volume->ondisk->vol0_next_tid < trans->tid)
684                         root_volume->ondisk->vol0_next_tid = trans->tid;
685                 hammer_crc_set_volume(root_volume->ondisk);
686                 hammer_modify_volume_done(root_volume);
687                 hammer_io_flush(&root_volume->io);
688         }
689
690         /*
691          * Wait for I/Os to complete
692          */
693         hammer_flusher_clean_loose_ios(hmp);
694         hammer_io_wait_all(hmp, "hmrfl2");
695
696         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
697                 goto failed;
698
699         /*
700          * Flush meta-data.  The meta-data will be undone if we crash
701          * so we can safely flush it asynchronously.
702          *
703          * Repeated catchups will wind up flushing this update's meta-data
704          * and the UNDO buffers for the next update simultaniously.  This
705          * is ok.
706          */
707         count = 0;
708         while ((io = TAILQ_FIRST(&hmp->meta_list)) != NULL) {
709                 if (io->ioerror)
710                         break;
711                 KKASSERT(io->modify_refs == 0);
712                 if (io->lock.refs == 0)
713                         ++hammer_count_refedbufs;
714                 hammer_ref(&io->lock);
715                 KKASSERT(io->type != HAMMER_STRUCTURE_VOLUME);
716                 hammer_io_flush(io);
717                 hammer_rel_buffer((hammer_buffer_t)io, 0);
718                 ++count;
719         }
720
721         /*
722          * If this is the final finalization for the flush group set
723          * up for the next sequence by setting a new first_offset in
724          * our cached blockmap and clearing the undo history.
725          *
726          * Even though we have updated our cached first_offset, the on-disk
727          * first_offset still governs available-undo-space calculations.
728          */
729         if (final) {
730                 cundomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
731                 if (cundomap->first_offset == cundomap->next_offset) {
732                         hmp->hflags &= ~HMNT_UNDO_DIRTY;
733                 } else {
734                         cundomap->first_offset = cundomap->next_offset;
735                         hmp->hflags |= HMNT_UNDO_DIRTY;
736                 }
737                 hammer_clear_undo_history(hmp);
738
739                 /*
740                  * Flush tid sequencing.  flush_tid1 is fully synchronized,
741                  * meaning a crash will not roll it back.  flush_tid2 has
742                  * been written out asynchronously and a crash will roll
743                  * it back.  flush_tid1 is used for all mirroring masters.
744                  */
745                 if (hmp->flush_tid1 != hmp->flush_tid2) {
746                         hmp->flush_tid1 = hmp->flush_tid2;
747                         wakeup(&hmp->flush_tid1);
748                 }
749                 hmp->flush_tid2 = trans->tid;
750         }
751
752         /*
753          * Cleanup.  Report any critical errors.
754          */
755 failed:
756         hammer_sync_unlock(trans);
757
758         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) {
759                 kprintf("HAMMER(%s): Critical write error during flush, "
760                         "refusing to sync UNDO FIFO\n",
761                         root_volume->ondisk->vol_name);
762         }
763
764 done:
765         hammer_unlock(&hmp->flusher.finalize_lock);
766
767         if (--hmp->flusher.finalize_want == 0)
768                 wakeup(&hmp->flusher.finalize_want);
769         hammer_stats_commits += final;
770 }
771
772 /*
773  * Return non-zero if too many dirty meta-data buffers have built up.
774  *
775  * Since we cannot allow such buffers to flush until we have dealt with
776  * the UNDOs, we risk deadlocking the kernel's buffer cache.
777  */
778 int
779 hammer_flusher_meta_limit(hammer_mount_t hmp)
780 {
781         if (hmp->locked_dirty_space + hmp->io_running_space >
782             hammer_limit_dirtybufspace) {
783                 return(1);
784         }
785         return(0);
786 }
787
788 /*
789  * Return non-zero if too many dirty meta-data buffers have built up.
790  *
791  * This version is used by background operations (mirror, prune, reblock)
792  * to leave room for foreground operations.
793  */
794 int
795 hammer_flusher_meta_halflimit(hammer_mount_t hmp)
796 {
797         if (hmp->locked_dirty_space + hmp->io_running_space >
798             hammer_limit_dirtybufspace / 2) {
799                 return(1);
800         }
801         return(0);
802 }
803
804 /*
805  * Return non-zero if the flusher still has something to flush.
806  */
807 int
808 hammer_flusher_haswork(hammer_mount_t hmp)
809 {
810         if (hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR)
811                 return(0);
812         if (TAILQ_FIRST(&hmp->flush_group_list) ||      /* dirty inodes */
813             TAILQ_FIRST(&hmp->volu_list) ||             /* dirty bufffers */
814             TAILQ_FIRST(&hmp->undo_list) ||
815             TAILQ_FIRST(&hmp->data_list) ||
816             TAILQ_FIRST(&hmp->meta_list) ||
817             (hmp->hflags & HMNT_UNDO_DIRTY)             /* UNDO FIFO sync */
818         ) {
819                 return(1);
820         }
821         return(0);
822 }
823